Pre-training sets low-rank geometry, RL fine-tunes shape
yunta_tsai · x · 2026-08-21
yuntatsai argues that pre-training establishes the correct low-rank geometry, while RL fine-tuning adjusts the shape. Better pre-training geometry requires less fine-tuning. Fine-tuning cannot fix low-rank mistakes because the optimizer cannot determine the correct shape if the observations have significant gaps.
Related event: Pretraining sets the low-rank geometry; RL only reshapes it(2 posts)→
More from Research
- poolside researchers on turning tens of thousands of data-mix experiments into frontier-moving models — arena · 2026-08-21
- poolside researchers on turning tens of thousands of data-mix experiments into frontier models — arena · 2026-08-21
- Google open-sources TIPSv2 image-text encoders, SOTA on all four zero-shot segmentation benchmarks — bdsqlsz · 2026-08-21
- Open-source WBC-Mjlab lands on mjswan Cloud: one policy drives G1's whole-body skills in your browser — kevin_zakka · 2026-08-21
- Watermarking research trivializes lexical and semantic choices, researcher says — TuhinChakr · 2026-08-21
- Time to integrate AI checks and rethink paper publication norms, economist argues — joshgans · 2026-08-21